Insights
Notes from ongoing measurement of AI recommendation behaviour in real estate markets. Written for operators, not for search engines.
Recommendation answers are synthesized, not ranked. Understanding the difference between retrieval, source trust and entity confidence explains most of what teams find surprising in their first audit.
Full note coming soon
Production volume is largely invisible to a language model. What is visible is what has been written, structured and corroborated about you in places the model can reach.
Full note coming soon
Across illustrative measurement, recommendation answers lean disproportionately on a small set of third-party sources per market. Coverage on those sources is unevenly distributed between competitors.
Full note coming soon
Single-run testing is noise. Locked prompt sets, repeated runs and captured timestamps are what turn AI answers into something you can actually measure a change against.
Full note coming soon
“Luxury realtor NYC” and “who should I hire to sell my $5M Tribeca apartment” are different problems. Only one of them ends with a name.
Full note coming soon
No one outside the labs knows the ranking systems. Anyone selling certainty about them is selling something else. Here is where our confidence ends.
Full note coming soon